Architecture synthesis for application specified parallel computers

研究了一种通过分析应用程序来构造静态可重构处理器间网络连接拓扑的方法。本文提出了一种处理单元模型的网络和并行程序生成算法。该算法被实现为一个网络生成系统，本研究开发的网络生成系统接收一个用C语言编写的应用程序、多个PE和一个PE的多个连接端口。接收到的应用程序被划分为操作，并且操作被分配给PE并被调度。PE间通信链路是基于数据交换量和执行时间生成的。调度过程重复两次，以细化生成的网络和任务分配。最后，系统生成一个PE间的连接信息和一个可在并行计算机上执行的并行程序，并通过网络生成系统生成网络和并行程序，并给出了一些示例程序。在并行计算机模拟器上对这些网络和程序进行了评估，结果表明，在生成的网络上，采用SOR格式的泊松求解器比传统的规则网络（如最近邻连接）运行得更快。FFT程序的另一个结果显示了使用生成系统的类似好处。这些实验也表明，本文所导出的系统对于具有非平衡结构的并行程序是更有效的。

A method is studied for composing the connection topology of statically reconfigurable inter-processor networks by analyzing application programs. An algorithm to generate networks and parallel programs for a PE (Processing Element) model is presented. This algorithm is realized as a network generation system.The network generation system developed by this research receives an application program written in C language, a number of PE, and a number of connection ports a PE. The received application program is divided into operations and the operations are assigned to PEs and scheduled. Inter-PE communication links are generated based on the amount of data exchange and execution time. Scheduling process in repeated twice to refine the generated network and task assignment. Finally, the system produces an inter-PE link information and a paralleled program to be executed on the parallel computer with the generated inter-PE link.By the network generation system, networks and paralleled programs are generated with some sample programs. These networks and programs are evaluated on a parallel computer simulator.As a result, the Poisson solver with SOR scheme runs faster on the generated network than conventional regular networks such as the nearest neighbor connection. Another result for the FFT program shows similar benefits using the generation system. These experiments also show the system derived here is more efficient for the case of a paralleled program with unbalance structure.